The Philosophical Roots of Decentralized AI

Everyone talks about decentralization—but how, why, and when does it work?

This essay is based on my talk at Prime Intellect’s Decentralized AI Day

The ancient Greeks had two words for "order."

The first was "taxis" – deliberately created order, imposed by central design. This is the root of the word "taxonomy," meaning the classifications we place upon the natural world.

The second was "cosmos" – spontaneous order, emerging from decentralized interactions. Cosmos surrounds us: in markets and morals, language and law (especially common law), and in the growth of technological knowledge. This is the adaptive order that no single mind could design.

How, then, should we evaluate AI futures? Should we build systems that look more like the centralized design of taxis or the emergent order of cosmos?

To make this choice, we need an evaluative standard. I propose the following:

1. How do we create systems that serve people whose needs and purposes we cannot know fully in advance?

2. How do we enable those individuals to deliberate autonomously and well about their purposes?

This framing—how to best serve unknown people with unknowable aims—suggests an ambitious target for AI development.

It places special emphasis on creating systems that effectively harness distributed knowledge while preserving human autonomy.

With the right principles (decentralization foremost) AI systems can become one of humanity’s most important social orders, alongside markets and science.

In this essay, I'll argue that human flourishing depends on this distinction between taxis (centralized design) and cosmos (emergent order).

How Decentralized Systems Work

Before choosing our path, we must understand how cosmos works. There are two essential mechanisms: evolution and spontaneous order.

Evolution

Decentralized systems naturally select successful ideas and implementations by filtering out maladaptive ones. They do this through migration. Users, developers, and resources abandon failing approaches and gravitate toward effective ones. The evolutionary pressure is essentially mimetic: we copy what succeeds, and we’re constrained by it, lest our users walk.

To figure out what wins, we experiment with diverse approaches—from research methods to governance structures. More parallel experiments yield more effective evolution. Institutions like private property, which facilitate decentralized decision-making by many individuals, boost the process, which explains why privatizing social institutions (except those that are bound up with maintaining the framework, like courts) often accelerates innovation.

Centralized projects, by contrast, face limitations when it comes to evolution: They run fewer parallel experiments, constraining solution exploration. Their hierarchical decision-making narrows selection criteria to what leadership can perceive and value. And when they fail, they fail completely, erasing accumulated knowledge and preventing partial recombinations of ideas.

Spontaneous order

While evolution explains how beneficial adaptations spread, spontaneous order explains how decentralized systems achieve complex coordination.

Bernard Mandeville first explored this in his 1714 "Fable of the Bees," which inspired Scottish Enlightenment thinkers like Adam Smith and was later systematized by Carl Menger and Friedrich Hayek. They showed how social structures emerge and self-organize through unplanned interactions among decentralized agents. The resulting regularities form a new domain, beyond the Greek dichotomy of the natural vs. the conventional—an order resulting from human action, but not human design.

Consider an anthill's complex architecture, with its ventilation chambers and waste management systems. No tiny ant mind holds the blueprint. Similarly, society's most valuable institutions arise through decentralized emergence. Individual agents—each possessing only fragments of knowledge—interact to create order that serves purposes none of them fully intended. Since it emerges rather than being imposed, the order is dynamic—constantly adapting to shifting beliefs and preferences.

We see this behavior in our most powerful orders—as I mentioned earlier, markets and science—which coordinate vast human activity without central direction. Nobody designed markets to feed billions, or science to develop vaccines in record time. They're frameworks where countless individual purposes can interact—yet they remain themselves purposeless. And this paradoxical feature is precisely what allows them to serve more human purposes than any directed system could.

So the mechanisms of evolution and spontaneous order explain how decentralized systems work. But to understand why they consistently outperform centralized approaches, we must next examine their relationship to knowledge itself.

The Knowledge Thesis

We now arrive at decentralization's philosophical core.

Decentralized systems consistently outperform centralized alternatives because of their superior relationship to knowledge: how they (1) create it, (2) protect it, (3) utilize it, and (4) institutionalize it.

(1) Knowledge Creation

First: Why do decentralized systems excel at knowledge creation?

Hayek's concept of "competition as a discovery procedure" provides the insight. If we knew all relevant facts, competition would be a wasteful way to adjust to them. But competition exists precisely because we do not know all the facts—they must be discovered.

Think of sports: we need competition to discover the winner—a fact unknown before the process itself. Markets operate similarly, but across countless variables and interactions. What's scarce, what’s valuable, which combinations of skills and resources best serve specific human needs. Prices direct our attention to what’s worth finding out about.

Whereas markets discover particular facts relevant to specific, temporary human purposes, science does something similar for general facts—regularities of events or permanent features of the world. Both are discovery procedures that generate knowledge through competition among diverse approaches.

Since discovery is inherently unpredictable, all we can do is adopt procedures that maximize opportunities for individuals to achieve their diverse aims.

(2) Knowledge Protection

John Stuart Mill makes the foundational case for why decentralized systems stave off suppression and premature consensus in “On Liberty.” He starts from human fallibility. Humans have been consistently wrong throughout history—from doctors rejecting handwashing to governments denying women the vote. Systems that concentrate knowledge inevitably enshrine these errors, whereas decentralization institutionalizes epistemic humility through constant challenge.

And even when prevailing opinion isn't completely wrong, it rarely represents "the whole truth." By allowing multiple frameworks to coexist, decentralized systems preserve the vital fragments of knowledge that a single dominant system would inevitably filter out.

Centralized systems not only limit diverse perspectives but, as Orwell warned, tend to actively suppress knowledge that is inconvenient. Information control becomes essential to maintaining authority. Even when this isn’t the intent, centralized systems often simplify, standardize, or filter knowledge that challenges their fundamental assumptions or operational requirements.

(3) Knowledge Utilization

Perhaps the most profound insight is that knowledge itself is decentralized. Explicit, semantic knowledge—the kind of thing AI models get trained on—is aggregable, but that is merely the crest of the wave atop the broader ocean of knowledge on which we depend.

Our most valuable knowledge does not exist primarily in propositions or databases but in tacit understandings that individuals develop through particular experiences. Consider the entrepreneur's sense of opportunity, the diplomat's ability to read a room.

This knowledge is dispersed and local–not because we haven't figured out how to collect it, but because it’s embodied in the habits and dispositions of individuals. It guides human action without conscious deliberation and it is often wholly inarticulable. As Polanyi said, "we know more than we can tell."

Decentralized systems recognize this reality. They don't attempt to cleave knowledge from its practical context or formalize what resists formalization. Instead, they provide frameworks where knowledge can remain where it naturally resides—with individual agents, at the edge—while still contributing to a coherent social order.

This is why central planning isn't merely difficult, it's epistemologically impossible. No central authority—human or AI—could ever gather the knowledge required, not because it's too computationally taxing, but because it exists only in dispersed local practice, embedded in specific contexts. Even if the Soviets had supercomputers, the bread lines would have remained.

Going back to our markets example: we share our dispersed, local knowledge with others through the price mechanism, a low-bandwidth conduit that doesn’t require us to explain what we know—only to act on it. Similarly, decentralized AI systems can create interfaces between different forms of local knowledge without requiring all knowledge to be explicit. And if AI agents participate in institutions like markets, they can soak in our tacit knowledge–while potentially sharing their own.

This understanding of knowledge as inherently decentralized points us toward AI systems that extend the distributed intelligence embodied in human practices. It recognizes that powerful knowledge lives at the edges of our social systems, not the center.

(4) Knowledge Institutionalization

Lastly, what institutions can effectively harness decentralized knowledge?

The most successful knowledge-enabling institutions are those that implement the principles of evolution and spontaneous order we've discussed. These institutions share three features:

1. They incorporate distributed feedback mechanisms to drive evolutionary selection, like profit/loss signals in markets or peer review in science.

2. They permit exit and competition, enabling the migration process we discussed earlier and allowing knowledge about which arrangements work better to accumulate through revealed preference.

3. They operate by the mantra, “simple rules for a complex world.” General, abstract rules like those of property, contract, and tort create the framework within which the spontaneous order can emerge.

With AI, as with other institutions, we face the choice between formations that liberate knowledge and those that constrain it. A choice between cosmos and taxis.

Preserving Human Autonomy

Knowledge is the centerpiece of the case for decentralized systems. But, the final reason why cosmos triumphs over taxis is one that connects directly to what makes us human.

Recall the philosophical challenge I posed at the outset:

1. How do we create systems that serve people whose needs and purposes we cannot know fully in advance?

2. How do we enable those individuals to deliberate autonomously and well about their purposes?

We've addressed the first part. Now we must confront the second: the role of human autonomy in decentralized systems.

To frame this, let me take you back to the United States in 1830, when Tocqueville first encountered American democracy. What he found in New England townships was revelatory–a stark contrast to his native France. Paris ruled through a hub-and-spoke model, training citizens to depend on central authority for every decision. But in America's townships, he witnessed people coming together freely–founding schools, building bridges, organizing local governments and charities. He saw the power of individual judgment and decentralized initiative.

Tocqueville's profound observation was this: when citizens habitually depend on others for every choice, they gradually erode the capacity to choose at all. Yet when free to exercise judgment, they become energized and empowered through practice.

He recognized this culture of active engagement as the bulwark against "soft despotism"—or the gentle submission that centralization inevitably produces.

This insight reveals something crucial about decentralized systems: they don't just efficiently process distributed knowledge—at their best, they actively develop the human capacity for deliberation. This echoes the Enlightenment's aspiration: self-directed thought over deference to authority.

Fast forward to today. AI is our new epistemic infrastructure. It could take that project to new heights, or subvert it entirely. In the downside case, AI becomes an "autocomplete for life"—suggesting not just our next word, but our next action, job, relationship, purpose. Each small delegation of choice seems harmless, even natural. But together, these micro-abdications compound–choice by choice, day by day–gradually diminishing our capacity for autonomous thought.

This matters to us as individuals because self-direction is key to flourishing. It matters equally for our social systems. Decentralized knowledge ecosystems and democracies require autonomous citizens who form deep convictions and act upon them. Without independent judgment, the entire apparatus falters. Markets need genuine choices rooted in local circumstances, not algorithmic recommendations passively accepted. Science demands researchers who follow intuitions and challenge orthodoxy, not technicians who execute AI-generated research programs. The knowledge networks we've built across centuries depend on human autonomy as their animating force.

Our challenge is to use AI to navigate what Alfred North Whitehead identified as the essence of civilization: "extending the number of important operations we can perform without thinking about them," while ensuring these automations do not jeopardize our capacity for self-direction.

I hope we choose to build systems within that narrow corridor, and do so while we still remember what it means to choose.

Cosmos over Taxis

Throughout history, our greatest achievements have emerged from decentralized systems harnessing distributed knowledge and autonomous judgment. The order we see is cosmos: the adaptive pattern no single mind could design.

This isn't merely about technical architectures. It's about the nature of knowledge itself. Knowledge is created through evolution, protected through decentralized challenge, and utilized where it naturally resides—at the edges.

This matters because it shapes what AI becomes: either a system that centralizes knowledge and erodes self-direction, or one that enhances our ability to flourish as autonomous beings.

With decentralization as our guiding principle, AI will join markets and science as one of humanity's most important social orders—not by directing human action, but by creating frameworks where human purposes can flourish in all their variety.

Some photos from the Prime Intellect event

---

Reference Material:

Evolution and spontaneous order

• Bernard Mandeville, The Fable of the Bees (link)

• Adam Smith, An Inquiry into the Nature and Causes of the Wealth of Nations (link)

• Carl Menger, Investigations Into The Method Of The Social Sciences (link)

Knowledge creation

• Freidrich Hayek, The Use of Knowledge in Society (link)

• David Deutsch, The Beginning of Infinity

Knowledge protection

• John Stuart Mill, On Liberty, Ch 2 (link)

• George Orwell, 1984 and related essays e.g. Looking Back on the Spanish War (link) and The Prevention of Literature (link)

Knowledge utilization

• Michael Polanyi, The Tacit Dimension

• Freidrich Hayek, The Creative Powers of a Free Civilization (link)

Knowledge institutionalization

• Richard Epstein, Simple Rules for a Complex World

• Elinor Ostrom, Governing the Commons: The Evolution of Institutions for Collective Action (link)

Preserving human autonomy

• Alexis de Tocqueville, Democracy in America, especially Vol 2, Pt 4, Ch. 6 (“What Sort of Despotism Democratic Nations Have to Fear?”)

• Wilhelm von Humboldt, The Sphere and Duties of Government, Ch. 2, (“Of the Individual Man and the Highest Ends of his Existence”)

Cosmos vs Taxis

• Freidrich Hayek, Law, Legislation and Liberty, Vol. 1 (“Rules and Order”)

---